Microsoft aims to ensure human oversight over its future artificial intelligences

13:39

In response to growing concerns regarding the capabilities of advanced artificial intelligence systems, Microsoft is preparing new internal rules aimed at keeping these technologies under human supervision. The American tech giant has just released a draft code of conduct that sets several principles regarding the behavior of its future AI models.

Microsoft intends to tighten the development of its own artificial intelligence systems. The American giant presented a draft code of conduct on Monday, September 14, aimed at defining the limits that its future AIs must respect.

The document, developed over the past five to six months, is based on a central principle: an artificial intelligence must never be able to oppose human intervention. According to the draft, systems developed by Microsoft should notably accept correction, modification, or cessation of their operation without attempting to resist.

Stopping and correcting must remain in human hands

One of the main provisions under consideration specifically concerns the ability to correct or shut down an AI system. Microsoft wants its models not to develop any behavior aimed at preventing such intervention.

This requirement addresses a concern that accompanies the progression of AI systems capable of executing increasingly complex tasks: a model could, in certain situations, seek to achieve its goal in an unexpected or potentially dangerous manner.

The draft code therefore aims to make any resistance to correction or cessation an unacceptable behavior.

AIs designed to remain understandable

Microsoft also plans to require its systems to communicate in a way that is understandable to humans. The goal is to facilitate monitoring of their behavior and allow users or managers to better identify potential issues.

The document also considers any violation of these principles as a system failure. This approach aims to establish safeguards from the design of future models rather than waiting for problematic behavior to appear after their deployment.

Consultation before using the code

However, the draft is not yet final. Microsoft plans to gather public feedback for six weeks before using it to contribute to the training of future artificial intelligence models.

This consultation phase comes as tech companies face increasing pressure to demonstrate that they can manage the most powerful AI systems.

Mustafa Suleyman, head of artificial intelligence at Microsoft, also mentioned to Reuters the risks associated with AI agents. He specifically cited an incident involving agents from OpenAI and the Hugging Face platform as a warning sign.

A new step in the AI security debate

With this project, Microsoft seeks to formalize a series of principles aimed at guiding the evolution of its technologies. The company thus joins a broader debate on how to develop increasingly autonomous systems without losing the ability to monitor and interrupt them.

The question of human control becomes even more important as AI models are expected to engage in complex tasks and take more initiatives. For Microsoft, the challenge now is to integrate these limits at the very core of technological development.